Method and System for Providing Multilingual Menu Translation by Combining Knowledge Data and Language Models

Abstract

A method for providing multilingual menu translation by fusing knowledge data and a language model according to one technical aspect of the present invention is performed in a multilingual menu translation system and comprises: receiving a user request including original menu information to be translated; searching for domain information related to the menu information in knowledge data based on the original menu information; generating a prompt for translating the original menu information based on the searched knowledge data; and inputting the generated prompt into a large-scale language model to generate a multilingual translation result.

Description

Technology Field

[0001] The present invention relates to multilingual translation technology, and more specifically, to a method and system for providing multilingual menu translation that fuses knowledge data and a language model regarding information related to food menus to provide high-quality translation and automatically provide various information about the menu. Background Technology

[0002] As global food cultures converge and the global foodservice industry expands, the demand for multilingual translation of food menus is continuously increasing. In particular, providing accurate and natural-sounding food information to users of various languages ​​has become an essential element in tourist destinations, multinational franchises, public institutions, and online delivery platforms.

[0003] Existing multilingual translation technologies have primarily evolved by relying on statistical machine translation (SMT) or neural network-based translation (NMT), but the following limitations exist when it comes to food menu translation.

[0004] In conventional methods, first, food names, cooking methods, and ingredient names are often expressed using regional dialects, loanwords, or abbreviations, making it difficult for standard translation models to accurately grasp their meaning. Second, even for the same food, names and cooking methods vary by region or cultural sphere, which can lead to user confusion if translated literally. Third, existing translation technology is limited to translating only menu names and fails to provide information crucial for user selection, such as main ingredients, allergens, and spiciness levels. Prior art literature

[0005] Korean Registered Patent Publication No. 10-2732291 (Publication Date: Nov. 15, 2024) The problem to be solved

[0006] One technical aspect of the present invention is to provide a method and system for providing multilingual menu translation by fusing knowledge data and a language model, which can translate the names of food menus more accurately and naturally by fusing knowledge data and a language model, thereby solving the problem that conventional general translation systems have difficulty conveying accurate meaning regarding specialized domains such as food menus and cause contextual misunderstandings regarding dialects or regional food names.

[0007] Furthermore, one technical aspect of the present invention is to provide a method and system for providing multilingual menu translations by fusing knowledge data and a language model, which can provide key supplementary information such as main ingredient information, allergy-causing ingredients, and spiciness levels along with multilingual translation results for food menus, thereby solving the problem that existing translation systems only translate menu names and fail to provide information on the composition or ingredients of food.

[0008] Furthermore, one technical aspect of the present invention is to provide a method and system for providing multilingual menu translations by fusing knowledge data and a language model, which can improve the reliability and consistency of sentence generation through a Retrieval-Augmented Generation (RAG) structure that searches for and extracts external knowledge data to provide contextual information to the language model in advance, thereby solving the problem where large-scale language models have strengths in general sentence generation but lack domain-specific knowledge, resulting in low accuracy in food-related translations.

[0009] In addition, one technical aspect of the present invention is to provide a method and system for providing multilingual menu translation by fusing knowledge data and a language model, which not only improves translation accuracy but also automatically processes and outputs customized information by reflecting the user's language settings, preferences, allergy conditions, etc., and provides an automatic translation service infrastructure applicable to actual operations through a structure that can be linked to various service environments such as food service companies or tourism information platforms.

[0010] The above-mentioned objectives and various advantages of the present invention will become more apparent to those skilled in the art from the preferred embodiments of the present invention. means of solving the problem

[0011] One technical aspect of the present invention proposes a method for providing multilingual menu translation by fusing knowledge data and a language model. The method for providing multilingual menu translation by fusing knowledge data and a language model is performed in a multilingual menu translation system and includes: a step of confirming a user request including original menu information to be translated; a step of searching for domain information related to the menu information in knowledge data based on the original menu information; a step of generating a prompt for translating the original menu information based on the searched knowledge data; and a step of inputting the generated prompt into a large-scale language model to generate a multilingual translation result.

[0012] In one embodiment, the step of searching for domain information related to the menu information in knowledge data may include extracting keywords based on food names included in the original menu information and searching for related information containing the keywords in the knowledge data based on content.

[0013] In one embodiment, the knowledge data comprises a plurality of entities classified into a cooking entity corresponding to a food name, an ingredient entity corresponding to an ingredient, and an origin entity corresponding to a region; and a knowledge graph expressing the semantic relationships between the plurality of entities in a triple structure; wherein the knowledge graph can provide the association information for translation by exploring a related relationship path based on the entity associated with the extracted keyword.

[0014] In one embodiment, the step of searching for association information including the keyword in knowledge data based on content may include: a step of expanding the search for entities including expressions or synonyms similar to the extracted keyword; a step of evaluating multiple paths based on relationship attributes between the expanded entities and other entities; and a step of setting the association information by preferentially selecting a path with high association.

[0015] In one embodiment, the step of setting the association information by preferentially selecting the highly associated path may include the step of setting the path containing the material object or region object most closely connected to the food name as the priority evaluation target.

[0016] In one embodiment, the step of generating the prompt may include: generating a food description sentence by reflecting at least one of a food name, ingredients, a recipe, and regional characteristics, including object information extracted from the searched domain information; and generating a structured prompt including multilingual translation target language, translation tone, and stylistic information together with the food description sentence.

[0017] In one embodiment, the step of generating the multilingual translation result may include: a step of identifying a plurality of translation candidates generated from a language model based on the prompt; and a step of calculating the semantic similarity between the food name to be translated and the knowledge data for each translation candidate, and selecting the translation candidate with the highest semantic similarity as the final translation result.

[0018] In one embodiment, the step of generating the multilingual translation result may include: determining user characteristics based on the user's language settings, location information, or terminal information from the user terminal; and dynamically setting the translation style, explanation level, and output item according to the user characteristics to generate the translation result.

[0019] Another technical aspect of the present invention proposes a multilingual menu translation system that fuses knowledge data and a language model. The multilingual menu translation system comprises: a seller terminal that provides a seller request including original menu information to be translated; and a service server that, based on the original menu information, searches for domain information related to the menu information in knowledge data, generates a prompt for translating the original menu information based on the searched knowledge data, and inputs the generated prompt into a large-scale language model to generate a multilingual translation result.

[0020] The means for solving the above-mentioned problem do not enumerate all features of the present invention. Various means for solving the problem of the present invention may be understood in more detail by referring to specific embodiments in the following detailed description. Effects of the invention

[0021] According to one embodiment of the present invention, by translating food menu information by fusing knowledge data and a language model, it is possible to generate translation results that maintain high accuracy and natural context even for expressions that are difficult for general translation systems to process, such as dialects, loanwords, and regional special expressions, thereby having the effect of improving user satisfaction.

[0022] In addition, according to one embodiment of the present invention, by not only translating the names of food menus but also providing key additional information such as main ingredient information, allergy-causing components, and spiciness levels for the food, it is possible to support customized food selection for the user and provide information accessibility that takes into account conditions such as allergies or dietary restrictions.

[0023] In addition, according to one embodiment of the present invention, through a Retrieval-Augmented Generation (RAG) structure that searches for and extracts information based on external knowledge data and reflects it in the prompt configuration of a language model, the translation quality and consistency can be improved compared to the operation of a language model alone.

[0024] In addition, according to one embodiment of the present invention, by integrating practical service infrastructure such as multilingual translation functions, various user condition settings, integration with F&B companies, and linkage with tourism platforms, it is possible to secure scalability and commercial applicability applicable throughout the food service and tourism industries.

[0025] The effects of the invention described above are not limited to all the various effects of this application, and these various effects can be easily understood from the specific embodiments in the detailed description. Brief explanation of the drawing

[0026] FIG. 1 is a diagram illustrating an overview of the overall configuration of a multilingual menu information providing system according to one embodiment of the present invention. FIG. 2 is a flowchart illustrating the flow of a method for providing multilingual menu information according to an embodiment of the present invention. FIG. 3 is a configuration diagram illustrating the internal configuration of a service server (300) according to an embodiment of the present invention and the linkage relationship of an external configuration (large-scale language model (400), knowledge data (500)) linked thereto. FIG. 4 is a flowchart illustrating the linkage structure between a service server (300) and a large-scale language model (400) and a prompt processing procedure for generating food descriptions or additional information according to an embodiment of the present invention. FIG. 5 is a diagram illustrating an example of the configuration of a food ingredient knowledge graph according to one embodiment of the present invention, and the nodes and relationship structure included in the graph. FIG. 6 is a diagram illustrating a flowchart of knowledge graph-based information extraction and structuring processing according to one embodiment of the present invention. FIG. 7 is a flowchart illustrating a flow of generating a multilingual translation prompt using knowledge graph-based path information according to an embodiment of the present invention, and performing post-processing and expression adjustment according to user settings on the generated translation result. FIG. 8 is a block diagram illustrating the configuration of a multinational visitor-based restaurant information provision system according to one embodiment of the present invention. FIG. 9 is a flowchart illustrating a restaurant review processing procedure for estimating multinational visitor information according to an embodiment of the present invention. FIG. 10 is a diagram illustrating an example of the configuration of request data received when uploading a review, according to an embodiment of the present invention. FIG. 11 is a flowchart showing the processing sequence for estimating nationality or cultural sphere based on user language preference information according to one embodiment of the present invention. FIG. 12 is a block diagram illustrating the configuration of a web-based multilingual menu ordering system according to one embodiment of the present invention. FIG. 13 is a flowchart illustrating the processing procedure of a web-based multilingual menu ordering method capable of location-based order verification according to an embodiment of the present invention. FIG. 14 is a block diagram illustrating an order management unit that automatically analyzes abnormal order behavior of a user and presents correction candidates according to an embodiment of the present invention. Figure 15 is a flowchart illustrating an order management method performed in the order management department of Figure 14. FIG. 16 is a block diagram showing the configuration of a multilingual question-and-answer providing system using an intelligent chatbot according to one embodiment of the present invention. FIG. 17 is a flowchart showing the processing flow of a method for providing multilingual question and answer using an intelligent chatbot according to an embodiment of the present invention. FIG. 18 is a flowchart illustrating a method for generating interactive response data for a user query according to an embodiment of the present invention. FIG. 19 is a flowchart illustrating an expert pool classification and dynamic selection method for selecting appropriate HITL (Human-In-The-Loop) experts based on the reliability and characteristics of user queries according to an embodiment of the present invention. FIG. 20 is a diagram illustrating an exemplary computing operating environment of a service server according to an embodiment of the present invention. Specific details for implementing the invention

[0027] Preferred embodiments of the present invention will be described below with reference to the attached drawings.

[0028] However, embodiments of the present invention may be modified in various different forms, and the scope of the present invention is not limited to the embodiments described below. Furthermore, embodiments of the present invention are provided to more fully explain the present invention to those with average knowledge in the relevant technical field.

[0029] That is, the aforementioned objectives, features, and advantages will be described in detail below with reference to the attached drawings, and accordingly, a person skilled in the art to which the present invention pertains will be able to easily implement the technical concept of the present invention. In describing the present invention, detailed descriptions of known technologies related to the present invention are omitted if it is determined that such descriptions may unnecessarily obscure the essence of the present invention. Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the attached drawings. In the drawings, the same reference numerals are used to indicate the same or similar components.

[0030] Additionally, singular expressions used in this specification include plural expressions unless the context clearly indicates otherwise. In this application, terms such as "composed of" or "comprising" should not be interpreted as necessarily including all of the various components or steps described in the specification, and should be interpreted as meaning that some of the components or steps may not be included, or that additional components or steps may be included.

[0031] In addition, various components and their sub-components are described below to explain the system according to the present invention. These components and their sub-components may be implemented in various forms, such as hardware, software, or a combination thereof. For example, each element may be implemented as an electronic configuration to perform a corresponding function, or as software itself that is executable in an electronic system, or as a functional element of such software. Alternatively, it may be implemented as an electronic configuration and corresponding executable software.

[0032] The various techniques described herein may be implemented with hardware or software, or, where appropriate, with a combination of all of these. Terms such as "Unit," "Server," and "System" as used herein may likewise be treated as equivalent to computer-related entities, namely hardware, a combination of hardware and software, software, or software at runtime. Furthermore, each function executed in the system of the present invention may be configured as a module and may be recorded in a single physical memory or distributed among two or more memories and recording media.

[0033] Various flowcharts are disclosed to explain embodiments of the present invention, but these are for the convenience of explaining each step and do not necessarily mean that each step must be performed in the order of the flowchart. That is, each step in the flowchart may be performed simultaneously, in the order according to the flowchart, or in the reverse order of the flowchart.

[0035] FIG. 1 is a diagram illustrating an overview of the overall configuration of a multilingual menu information providing system according to one embodiment of the present invention.

[0036] The present system includes a seller terminal (100) that registers original menu information, a user terminal (200) that receives multilingual translation results, and a service server (300) that performs menu translation and information provision while transmitting and receiving information from the two terminals.

[0037] The seller terminal (100) is a device used by a store owner or restaurant manager and provides an interface for registering menu information. For example, the owner can input original menu information, such as a specific menu name, menu description, ingredients, cooking method, and regional origin, through the seller terminal (100). This information is transmitted to and stored by the service server (300).

[0038] The user terminal (200) is a device used by general users, such as foreign users visiting the store. The user can access the digital menu page of the store by scanning a QR code provided at the offline store.

[0039] For example, a user terminal (200) can select a desired target language (e.g., English, Japanese, Chinese, etc.) through a default language setting value or a UI. The selected language setting is transmitted from the user terminal (200) to the service server (300).

[0040] The service server (300) is a central processing unit that performs functions such as registering menu information, translating, generating additional information, and customizing user output. Although not shown in the drawing, the service server (300) can implement functions such as translating menu information, generating additional information, and customizing user output by fusing knowledge data (500, shown in FIG. 3) and a large-scale language model (400, shown in FIG. 3).

[0041] In one embodiment, the service server (300) includes a domain information search function. For example, the service server (300) can extract keywords from menu names included in a user request and search for related domain information through extended search and path evaluation of the keywords in knowledge data.

[0042] In one embodiment, the service server (300) includes a prompt generation function. For example, the service server (300) generates a structured prompt including a target language, translation tone, style, etc., based on menu body information or a draft of additional information description.

[0043] In one embodiment, the service server (300) includes a multilingual translation function. For example, the service server (300) inputs a generated prompt into a large-scale language model to generate a translation result, and if multiple candidates exist, evaluates semantic similarity to select the optimal translation.

[0044] In one embodiment, the service server (300) includes a function for generating additional information. For example, the service server (300) extracts relevant information such as material objects, allergy objects, and spiciness objects from knowledge data and generates a draft of additional information description composed of the information and explanatory sentences.

[0045] In one embodiment, the service server (300) includes a user-customized optimization function. For example, the service server (300) dynamically adjusts the translation style or explanation level based on language settings, location information, terminal information, etc. received from the user terminal (200) to provide multilingual menu information in a manner most suitable for the user.

[0046] In one embodiment, the service server (300) includes an automatic sales information scraping function. For example, the service server (300) can identify the seller's online sales page (e.g., a store information page of a homepage or portal site) based on seller information, and crawl the menu information stored on the page to generate original menu information for each seller.

[0048] Hereinafter, with reference to FIGS. 2 to 10, various embodiments of a multilingual menu translation providing system and a method of providing the same, which fuse knowledge data and a language model, will be described.

[0050] FIG. 2 is a flowchart illustrating the flow of a method for providing multilingual menu information according to an embodiment of the present invention.

[0051] The method illustrated in FIG. 2 consists of the steps of collecting original menu information, requesting multilingual translation, processing translation, and providing results, centering on the interaction between a seller terminal (100), a user terminal (200), and a service server (300).

[0052] In step S201, the seller terminal (100) is a device used by a store manager or restaurant owner and provides original menu information written in a first language (e.g., Korean), such as food names, menu descriptions, ingredients, and cooking methods, to the service server (300).

[0053] In one embodiment, in addition to explicit registration through a seller terminal, the service server (300) can automatically check seller information and automatically collect (crawl) menu information from the seller's webpage, portal registration information, or online sales page (S202). In this case, the service server (300) has the advantage of being able to replace administrator input in that it automatically obtains original menu information written in the first language.

[0054] In step S203, the user terminal (200) is a device used by a user, such as a foreign user visiting the store. The user scans a QR code distributed within the store to access the digital menu page of the store, and in the process, selects a second language (e.g., English, Japanese, etc.) through the UI or default settings and sends a translation request. The request is transmitted to the service server (300).

[0055] In step S204, the service server (300) performs multilingual translation processing based on a second language notation request received from the user. At this time, the service server (300) may perform the following processing by fusing knowledge data and a large-scale language model, although not shown in the drawing.

[0056] First, the service server (300) performs multilingual translation of the original menu information written in the first language. In this process, the content of the menu body, such as the food name, cooking method, and regional origin of the menu, is reflected, and keyword-based domain information is searched from knowledge data and reflected in the prompt to minimize loss of meaning.

[0057] Additionally, the service server (300) constructs a draft of additional information description for the menu and performs additional multilingual translation thereon. Specifically, it can search for associations with material objects, allergy objects, spiciness objects, etc. within the knowledge data to generate an appropriate additional information sentence for the menu in natural language and convert it into a second language.

[0058] In step S205, the service server (300) provides the translated multilingual translation menu information to the user terminal (200). Here, the multilingual translation menu information refers to an integrated multilingual output result that includes not only the translation result of the original menu information, but also additional information such as descriptions of main ingredients, allergy information, and spiciness levels.

[0059] In one embodiment, the service server (300) can provide multilingual menu information customized to the user by adjusting the translation style, writing tone, and level of detail of the description, etc., taking into account the language settings, location information, and type of terminal of the user terminal (200).

[0060] In this way, the multilingual menu information provision method according to FIG. 2 can simultaneously improve the accuracy and satisfaction of food selection for foreign users by providing the menu text and additional information together through translation.

[0062] FIG. 3 is a configuration diagram illustrating the internal configuration of a service server (300) according to an embodiment of the present invention and the linkage relationship of an external configuration (large-scale language model (400), knowledge data (500)) linked thereto.

[0063] The service server (300) is a device that performs functions of collecting, storing, generating descriptions, multilingual translation, and user customization of menu information, and may include a sales information collection unit (310), a sales information management unit (320), a multilingual processing model (330), an explanation information generation AI (340), and an additional information generation AI (350).

[0064] The service server (300) can perform multilingual menu translation and information generation by linking with the large-scale language model (400) and knowledge data (500). In FIG. 3, the large-scale language model (400) and knowledge data (500) are described as being configured separately from the service server (300), but according to the embodiment, at least one function of the large-scale language model (400) and knowledge data (500) may be implemented as part of the function of the service server (300) or fused into the service server (300).

[0065] The sales information collection unit (310) performs the function of automatically or semi-automatically collecting menu information of a seller (e.g., a restaurant store manager).

[0066] In one embodiment, the sales information collection unit (310) may include an app-linked collector, a data scraper, and a multimodal processor.

[0067] The app-integrated collector can collect seller menu information in a structured manner through integration with restaurant POS (Point-of-Sales) systems, kiosk apps, order platform apps, etc. Alternatively, the app-integrated collector can collect original menu information from the seller's terminal through the seller's app by linking with the seller's app.

[0068] Meanwhile, the app-linked collector is not limited to specific collection targets based on its name; for example, it may collect information based on the web rather than an app. In other words, the app-linked collector is connected to the seller input interface of a service portal site, and can collect seller menu information when a seller inputs it through that interface.

[0069] Data scrapers collect information such as food names, prices, and descriptions from seller pages on the web—including portal sites, blogs, and delivery apps—using an unstructured crawling method, and perform HTML structure analysis and keyword filtering to configure menus and extract items. For example, on external seller pages such as Naver Place and Kakao Map or

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109]

[0110]

[0111] Prompt example

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118] Prompt example

[0119]

[0120]

[0121]

[0122]

[0123]

[0124]

[0125]

[0126]

[0127]

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0136]

[0137]

[0138]

[0139]

[0140]

[0141] Major Node Types

[0142]

[0143]

[0144]

[0145]

[0146] Major Relationship Types

[0147]

[0148]

[0149]

[0150]

[0151]

[0152]

[0153]

[0154]

[0155]

[0156] Prompt example

[0157]

[0158]

[0159]

[0160]

[0161]

[0162]

[0164]

[0165]

[0166]

[0167]

[0168]

[0169]

[0170]

[0171]

[0172]

[0173]

[0174]

[0175]

[0176]

[0177]

[0178]

[0179]

[0180]

[0181]

[0182]

[0183]

[0184]

[0185]

[0186]

[0187]

[0188]

[0189]

[0190] Prompt element

[0191]

[0192]

[0193]

[0194]

[0196]

[0197]

[0198]

[0199]

[0200]

[0201]

[0202]

[0203]

[0204]

[0205]

[0206]

[0207]

[0208]

[0209] Prompt configuration template

[0210]

[0211]

[0212]

[0213]

[0214]

[0215]

[0216]

[0217]

[0218]

[0219]

[0220]

[0221]

[0222]

[0223]

[0224]

[0225]

[0226]

[0227]

[0228]

[0229]

[0230]

[0231]

[0232]

[0233]

[0234]

[0235]

[0236]

[0237]

[0238]

[0239]

[0240]

[0242]

[0243]

[0244]

[0245]

[0246]

[0247]

[0248]

[0249]

[0250]

[0252]

[0253]

[0254]

[0255]

[0256]

[0257]

[0258]

[0259]

[0260]

[0261]

[0262]

[0264]

[0265]

[0266]

[0267]

[0268] User-Agent

[0269]

[0270]

[0271] Accept-Language field

[0272]

[0273]

[0274]

[0275]

[0276]

[0278]

[0279]

[0280]

[0281]

[0282]

[0283]

[0284]

[0285]

[0286]

[0287]

[0288]

[0289]

[0290]

[0291]

[0292]

[0293]

[0295]

[0296]

[0297]

[0298]

[0299]

[0300]

[0301]

[0302]

[0303]

[0304]

[0305]

[0306]

[0307]

[0308]

[0309]

[0311]

[0312]

[0313]

[0314]

[0315]

[0316]

[0317]

[0318]

[0319]

[0320]

[0321]

[0322]

[0323]

[0324]

[0325]

[0326]

[0328]

[0329]

[0330]

[0331]

[0332]

[0333]

[0334]

[0335]

[0336]

[0337]

[0338]

[0340]

[0341]

[0342]

[0343]

[0344]

[0345]

[0346]

[0347]

[0348]

[0349]

[0350]

[0351]

[0352]

[0354]

[0355]

[0356]

[0357]

[0358]

[0359]

[0360]

[0361]

[0362]

[0363]

[0364]

[0365]

[0366]

[0367]

[0368]

[0370]

[0371]

[0372]

[0373]

[0374]

[0375]

[0376]

[0377]

[0378]

[0379]

[0380]

[0381]

[0382]

[0383]

[0384]

[0385]

[0386]

[0387]

[0388] HITL Trigger Conditions

[0389]

[0390]

[0391]

[0392]

[0393]

[0394]

[0395]

[0397]

[0398]

[0399]

[0400]

[0401]

[0402]

[0403]

[0404]

[0405]

[0406]

[0407]

[0408]

[0409]

[0410]

[0411]

[0412]

[0413]

[0414]

[0415]

[0416]

[0417]

[0418]

[0419]

[0420]

[0421]

[0422]

[0423]

[0424]

[0425]

[0426]

[0427]

[0428]

[0429]

[0430]

[0432]

[0433]

[0434]

[0435]

[0436]

[0437] Expert Profile Information

[0438]

[0439]

[0440] Real-time availability filtering conditions

[0441]

[0442]

[0443]

[0444]

[0445]

[0446]

[0448]

[0449]

[0450]

[0451]

[0452]

[0453]

[0454]

[0455]

[0456]

[0457]

[0459] Information can be extracted by identifying menu blocks configured in a format. The multimodal processor performs multimodal processing on the collected seller menu information and can process image or voice-based menu information input. For example, the multimodal processor converts image or voice-based menu information into text-based menu information through OCR (Optical Character Recognition) and voice-to-text conversion technology. The sales information management unit (320) stores, refines, and processes the collected menu data to manage it in a structured format. The sales information management unit (320) can provide sales information templates, description information management, and additional information management functions. The sales information management unit (320) can manage sales information by configuring sales information templates. For example, the sales information management unit (320) can provide a structured template that automatically sorts the collected menu information by items such as food name, price, description, ingredients, cooking method, and regional characteristics. The sales information management unit (320) may include a description information management function. Here, the description information is structured and stored as detailed text description information regarding the menu description or main ingredients (e.g., "This is a menu featuring udon noodles in rich beef broth"), and can perform preprocessing tasks such as removing unnecessary phrases and grammar correction. The sales information management unit (320) may include an additional information management function. For example, the additional information may include at least one of the main ingredient information for the menu, allergy information for the menu set based on the main ingredient information, and spiciness information for the menu set based on the main ingredient information. Such additional information may be generated by an additional information generation AI (350), and the sales information management unit (320) can store and manage the generated additional information in association with each menu.The sales information management unit (320) manages additional information corresponding to the menu, such as allergy-causing ingredients, spiciness level, and regional origin, on an object basis, and this additional information is referenced in subsequent processing steps to create multilingual additional information descriptions. The multilingual processing model (330) is a module that performs a core function for improving translation quality in the present invention and can perform accurate translation by fusing a large-scale language model and knowledge data. The multilingual processing model (330) may include a prompt generation function and a dataset linkage function. The multilingual processing model (330) can search for domain information related to menu information in the knowledge data. For example, based on the knowledge data searched by the multilingual processing model (330), the additional information generation AI (350) can generate a draft of an additional information description for the menu. The prompts input to the large-scale language model (400) include domain-optimized examples (shots), object descriptions, translation instructions, etc. To this end, a multilingual dataset optimized for the corresponding menu domain can be combined based on a knowledge dataset stored in the knowledge data (500). For example, 'Kimchi stew' - 'Napa cabbage / Chili pepper / Pork / Korea' - Allergy and spiciness information - Expression methods by target language can be stored in association with the knowledge data. The multilingual processing model (330) can generate a prompt. For example, the multilingual processing model (330) can generate a structured prompt that includes a food description sentence along with the target language for multilingual translation, translation tone, and stylistic information. For another example, the multilingual processing model (330) can generate a structured prompt that includes the target language for multilingual translation, translation tone, and stylistic information along with additional information about the menu. For example, the multilingual processing model (330) can generate a structured prompt that reflects the target language for translation, stylistic tone (formal / informal), sentence length, output items (e.g., whether allergies are included), etc., based on the menu body information and additional information description.In one embodiment, a multilingual processing model (330) or a service server (300) may receive a plurality of translation candidates from a large-scale language model (400) and calculate the semantic similarity between the food name to be translated and the knowledge data for each translation candidate. The translation candidate having the highest value among the calculated semantic similarities may be selected as the final translation result. In one embodiment, the multilingual processing model (330) may dynamically change the prompt generation method based on the setting values ​​(e.g., language, location, terminal type, etc.) of the user terminal (200). For example, the multilingual processing model (330) may dynamically adjust the translation style, explanation level, and output items based on the language settings, location information, or terminal information of the user terminal (200). Through this, a user-customized multilingual translation result may be provided. The explanation information generation AI (340) is an artificial intelligence module that automatically generates natural language-based explanation sentences for the menu. The descriptive information generation AI (340) can perform combination of descriptive elements and generate sentences by reflecting contextual information such as ingredients, recipes, and regional origins in addition to the menu body. In one embodiment, the descriptive information generation AI (340) can generate descriptive sentences that highlight menu characteristics based on main ingredients, cooking methods, and origins (e.g., "It is a spicy stew with a gochujang base using Jeju pork"). The generated description can be used as a main sentence when generating a prompt before translation, and the writing style (formal / friendly, etc.), sentence length, and expression method can be adjusted as needed. The descriptive information generation AI (340) can be composed of a GPT-based pre-trained model or a finely tuned Transformer-family model, and fine-tuning specialized for the menu dataset is performed. The additional information generation AI (350) is an artificial intelligence that generates various additional information from menu information, such as allergy information, spiciness information, and precautions for each food ingredient.The additional information generating AI (350) can extract related entities from knowledge data (500) using menu names or ingredient names as keywords and infer whether they may cause allergies, the degree of spiciness, and the origin of the region. The additional information generating AI (350) can generate raw material information and allergy information. For example, for the "Dakgalbi" menu, the additional information generating AI (350) can generate allergy information based on the main ingredients after confirming "Main ingredients: chicken, gochujang." For example, after confirming "chicken - poultry protein" and "gochujang - soybeans, wheat" as raw materials, it can generate allergy information by checking whether each raw material is an allergen. The additional information generating AI (350) can also generate spiciness information. For the "Dakgalbi" menu, the additional information generating AI (350) can be set to "Spiciness: Medium or Higher" based on "Spiciness: Whether gochujang is included and the amount of chili peppers used." The additional information generating AI (350) can add a regional information description. For the "Dakgalbi" menu, a "regional information additional description" can be generated as "Regional origin: Chuncheon, Gangwon-do." As another example, regional information can also be used to verify raw materials. That is, for the "Yukgyejang" menu, beef (brisket, shank, round, etc.) is generally used, but for Jeollabuk-do and Jeollanam-do, pork (foreleg, loin, etc.) is used. Therefore, the regional information of the store can be checked, and the regional raw materials of "Yukgyejang" can be identified from the knowledge data to set the raw material information, etc. The large-scale language model (400) may be a general-purpose language model such as OpenAI GPT, Google PaLM, Meta LLaMA, etc., and in the embodiment of the present invention, RAG method fusion with the knowledge data (500) is applied to improve multilingual translation performance. For example, additional information is extracted through path search in a knowledge graph expressed as a triplet and constructed into an explanatory sentence. The knowledge data (500) is knowledge data that includes domain-specific multilingual expression information.For example, knowledge data (500) can be implemented as a knowledge graph. For example, the knowledge graph can be represented as a triplet structure of (entity-relationship-entity). For instance, it may include structures such as (kimchi stew, including, gochujang), (gochujang, including, soybean), (Jeolla-do, style, kimchi stew, including, pork), (Gyeongsang-do, style, kimchi stew, including, tuna). Through this, various related paths such as domain information, allergy information, regional origin, and ingredient descriptions related to a specific menu can be explored. Since this knowledge data is not in the form of a simple dictionary but is composed of relationship-centered network information, it is possible to generate multilingual prompts that reflect even the semantic context of the menu information. This is an advanced prompt optimization method in the application of a RAG (Retrieval-Augmented Generation) based language model, which is differentiated from existing simple LLM translation systems. FIG. 4 is a flowchart illustrating the linkage structure between a service server (300) and a large-scale language model (400) and a prompt processing procedure for generating food descriptions or additional information according to an embodiment of the present invention. This embodiment includes an overall flow in which, in order to accurately generate food description sentences or additional information such as allergy and spiciness information, the service server (300) configures a prompt, provides it to the large-scale language model (400), receives a response, and then stores and outputs the result. In step S401, the service server (300) performs input information settings for prompt generation. For example, if a menu name requiring translation (e.g., 'Dakgalbi') is provided by a user, and information on the main ingredients ('chicken, gochujang'), regional information ('Chuncheon'), and information on ingredients that may cause allergies exist for the menu, such information is collected as a single input information set. In addition, the user Language settings (e.g., English) or user characteristics (e.g., targeting elementary school students, request for an informal tone, etc.) received from the terminal (200) can also be collected at this stage.In step S402, the service server (300) determines and structures a prompt template based on the collected input information. For example, a 'prompt template for generating food description sentences' may be selected, in which case a prompt with the following structure may be generated: "Please write a natural food description sentence in English based on the following information. - Food name: Dakgalbi - Main ingredients: Chicken, Gochujang - Region: Chuncheon, Gangwon-do - Style: Friendly and simple sentences, understandable even to elementary school students." As another example, in the case of a prompt for extracting allergy information, it may be configured as follows: "Please tell me the components that may cause allergies regarding the following food ingredients. - Chicken, Gochujang - Output format: [Food ingredients] - [Allergy-causing items]" As such, the prompt configuration is templated and can be classified for explanatory or analytical purposes depending on the purpose. In step S403, the service server (300) calls a large-scale language model (400) and transmits the prompt structured above. In this embodiment, the large-scale language model (400) It may be composed of OpenAI GPT, Google PaLM, Meta LLaMA, etc., and the service server may request a prompt through an external API or an internal integration module. In step S404, the large-scale language model (400) generates a response based on the received prompt and sends it back to the service server (300). For example, in the case of generating a food description, the following response may be received: "Dak-galbi is a spicy stir-fried chicken dish made with gochujang sauce. It originates from Chuncheon, Korea, and is loved for its bold flavor." Or the allergy information response is as follows: "Chicken - poultry protein allergy," "Gochujang - soy allergy / wheat allergy." In step S405, the service server (300) refines the received response and performs format conversion according to the purpose.For example, post-processing such as style adjustment (formal / informal), simplification of expression (e.g., for infants), and sentence length limiting can be performed on food description responses. Additionally, in the case of allergy information, the response content can be converted to a JSON structure or field-based format so that it can be loaded into an internal storage. In step S406, the refined result is stored as a final translation result or supplementary information and can subsequently be output through a user terminal (200) or an administrator page. For example, supplementary information such as "Allergy: Soy, Wheat" and "Spiciness: Medium or higher" can be output along with the food description sentence in the multilingual translation menu information. This series of steps implements advanced language processing technology that goes beyond simple LLM API calls in that it includes an entity-level description through linkage with knowledge data (500) and is processed as a RAG (Retrieval-Augmented Generation) prompt configured based on the regional origin of each menu or the semantic association of the main ingredients. FIG. 5 shows an example of the configuration of a food ingredient knowledge graph according to an embodiment of the present invention and the nodes and relationship structures included in the graph. This is an urban diagram. This diagram is intended to explain the core operating principles of knowledge graph-based information inference and translation assistance by visually representing the relationships between nodes composed of food entities, ingredient entities, allergy entities, and region entities. The knowledge graph is structured in the form of a triplet (entity 1 - relationship - entity 2) and can represent various heterogeneous information by fusing it into a network structure. For example, the food "Gosari Yukgaejang" has a relationship in which it is a regional variant of "Yukgaejang" that originates from the "Jeju" region, can contain "Gosari" as a main ingredient, and can cause allergy components.The example in Figure 5 illustrates the configuration and application of such a knowledge graph, and the main node types and relationship types can be configured as follows: - Dish (Food): Yukgaejang, Bracken Yukgaejang, etc. - Ingredient (Ingredient): Bracken, Pork, Beef, etc. - Region (Region): Jeju, General, etc. - Allergy (Allergy): Soybeans, Crustaceans, Algae Protein, etc. - CONTAINS: Food contains a specific ingredient (e.g., Yukgaejang ―contains- Bracken) - ORIGINATED_FROM: Food or ingredient originates from a specific region (e.g., Bracken Yukgaejang ―originated_from- Jeju) - HAS_VARIANT: Regional variation of the food (e.g., Yukgaejang ―has_variant- Bracken Yukgaejang) - CAUSES_ALLERGY: Ingredients cause allergies (e.g., Bracken ―causes_allergy- Specific Allergy) Unlike a simple keyword dictionary, this knowledge graph [between] entities Since it is composed of a semantic network based on semantic and functional relationships, it is highly advantageous for generating context-based accurate information during the multilingual translation process. In the example of FIG. 5, an embodiment in which allergy information is included in the knowledge graph itself is illustrated, but it is not limited thereto, and allergy information can also be generated based on a large-scale language model as in the embodiment described in FIG. 4. For example, to explain the scenario of generating a menu description that reflects regional characteristics, if a user inputs a menu called 'Yukgaejang' and the region of the store is 'Jeju', regional variation information such as 'Gosari Yukgaejang' can be extracted through knowledge graph search. Accordingly, the service server (300) can generate a description such as "In Jeju, it is served as Gosari Yukgaejang containing Gosari," or include an expression that reflects the corresponding regional characteristics during translation.As another example, if we explain the scenario of automatic allergy information generation and translation, since the main ingredient of 'Beef Yukgaejang' is 'beef,' allergy information can be automatically generated by exploring connections such as 'beef ―causes_allergy- beef protein' through a knowledge graph. Furthermore, based on this, the following prompt for translation can be constructed: "Please write the allergy information in English based on the ingredients included in the following food. - Dish: Yukgaejang - Ingredients: Beef, Bracken - Output format: [ingredient] - [allergy risk]". In this way, a knowledge graph can infer domain-specific information that cannot be obtained from the food name itself through multi-path exploration between entities and information expansion based on semantic connections. By using such a knowledge graph, it is possible to secure both translation accuracy and information diversity by representing variations in ingredient composition or naming for the same food name depending on the region. Additionally, supplementary information such as allergies, regional characteristics, and recipes can be constructed through semantic relationships between nodes, rather than relying on single-keyword-based translations or descriptions. In addition, high-performance prompts can be constructed by selectively including only semantic-centered information through knowledge graph-based path search rather than directly injecting information into the language model, and path-based queries that are much faster and more flexible than relational databases are possible, which allows for fast and accurate processing of real-time dynamic search when generating prompts. Fig. 6 is a diagram illustrating a flowchart of knowledge graph-based information extraction and structuring processing according to an embodiment of the present invention. More specifically, a series of processes including synonym expansion, graph path search, path evaluation, and information construction are described based on keywords extracted from menu information (e.g., food name, descriptive sentence, etc.). Referring to Fig. 6, the service server (300) extracts keywords corresponding to core entities from the menu name or descriptive sentence (S610).For example, if a menu name such as "Jeju Bracken Yukgaejang" is entered, entity candidates such as 'bracken', 'Jeju', and 'Yukgaejang' are extracted. This keyword extraction can be performed by applying a predefined part-of-speech-based noun extraction or a deep learning-based NER (Named Entity Recognition) algorithm. Subsequently, the service server (300) performs a search for similar entities based on entity identifiers (aliases) within the knowledge data (500) based on the extracted keywords (S620). For example, the keyword 'bracken' can be expanded into the following expressions: 'bracken namul', 'bracken stem', 'bracken', 'bracken fern'. This can be achieved by configuring similar expression groups using the 'synonym', 'altLabel', and 'alias' attribute values ​​among the node attributes of the knowledge graph. In one embodiment, the service server (300) can select only semantically close keywords as expanded keywords by filtering them based on a Word2Vec-based similarity threshold (e.g., cosine similarity of 0.85 or higher). Subsequently, the service server (300) searches for a triplet relationship path in the form of (entity)-(relationship)-(entity) within the knowledge graph based on the expanded keywords (S630). For example, the following paths can be explored based on the entity 'Gosari': - ('Gosari', originated_from, 'Jeju') - ('Gosari', contained_in, 'Yukgaejang') - ('Gosari', causes_allergy, 'None') - ('Gosari', has_variant, 'Gosaritdae') Such path exploration can be implemented using BFS (Breadth-First Search) or weighted DFS, and is explored by reflecting the depth of the path (e.g., 2-hop or 3-hop limit), relationship type weights, node importance, etc.In particular, the present invention includes a relationship-based priority setting (e.g., 'originated_from' > 'contained_in' > 'has_variant') and a relationship-allowing filtering function based on object domain. Subsequently, the service server (300) selects a priority path from among the multiple paths found to be used for translation and additional information configuration (S640). To this end, scoring is performed for each path. Scoring may be based on the following elements: - Relationship weight (e.g., 'allergy-causing relationship' has high priority) - Path connection reliability (e.g., probability-based evaluation value when constructing a graph) - Path length (weights are assigned to shorter paths) - Entity importance (e.g., priority given to entities directly connected to menu names). At this time, the present invention may introduce a Multi-Criteria Decision Making (MCDM) algorithm for path evaluation, e.g., Analytic Hierarchy Process (AHP). The finally selected path is converted into base data for generating translation prompts or additional explanatory sentences. Subsequently, the service server (300) structures an input structure or explanatory sentence for multilingual translation based on the selected triplet path or node information (S650). For example, if the following path is selected: - ('Gosari', originated_from, 'Jeju') - ('Gosari', contained_in, 'Yukgaejang'), the following prompt elements may be generated: "Yukgaejang is a soup dish made using Jeju-produced bracken. Bracken is a local "It is a local specialty." Such structured sentences are transmitted to a prompt generation module (330) or a description generation AI (340) and used as input for a translation model. Additionally, the present invention may include a function that automatically converts an RDF (Triple-based) structure into a JSON-LD or GraphQL form and converts it into an input form optimized for LLM. This embodiment relates to an embodiment that independently generates food descriptions or allergy information based on a knowledge graph.However, this is not limited thereto, and as previously mentioned, at least some of the food descriptions or allergy information may be generated using a large-scale language model. FIG. 7 is a flowchart illustrating a flow of generating a multilingual translation prompt using knowledge graph-based path information according to an embodiment of the present invention, and performing post-processing and expression adjustment according to user settings on the generated translation result. Each step of FIG. 7 includes a series of processing steps in which the multi-path based structured information derived from FIG. 6 and user setting information are combined to construct a translation prompt optimized for the target language, the result is generated based on LLM, and then provided to a user terminal after translation quality evaluation and user-customized correction. In step S710, the service server (300) receives the structured domain information generated in step S650 of FIG. 6. For example, if the path extracted from the menu keyword 'Yukgaejang' is identified as - Yukgaejang - Included - Bracken - Yukgaejang - Regional Origin - Jeolla-do - Included - Pork - Yukgaejang - Regional Origin - Gyeongsang-do - Included - Beef, the corresponding path-based information The information is received by normalizing it into a single information structure (e.g., a tree structure or a multiple triple list). For example, the information structure at this time may be composed of a triple structure array in the form of (subject, relationship, object), or it may be composed in the form of a probabilistic adjacency matrix containing weight scores (frequency, importance, connectivity, etc.) for each node and edge during graph traversal. In step S720, the service server (300) receives the translation target language, stylistic tone, output item filtering settings, etc., received from the user terminal (200). For example, if the system language of the user terminal is Japanese and "formal sentences," "include allergy information," and "exclude map information" are specified in the user setting UI, the corresponding information is transmitted to the service server as a setting value in JSON format.In step S730, the service server (300) determines a prompt configuration template by reflecting the structured domain information received in S710 and the user settings of S720, and generates a translation prompt according to the template. For example, the prompt configuration template may be in the following format: [Food Name]: Yukgaejang [Main Ingredients]: Bracken, Beef [Regional Characteristics]: Gyeongsang-do [Allergy Information]: Beef Protein [Output Format]: Japanese, Formal, Sentence Length. The intermediate prompt template may be defined by matching an appropriate item among the domain-specific pre-trained templates pre-registered in the multilingual processing model (330) in the form of a conditional control statement (if / else), or by defining a configuration rule based on structured YAML or JSON. In step S740, the large-scale language model (400) receives the prompt from S730 and generates multiple translation candidates, and among these, the service server (300) may select the optimal translation based on the semantic similarity score. For example, the large-scale language model (400) [for the input prompt] Multiple translation results may be returned. For example, multiple candidates containing various writing styles and expressions may be generated as shown below. The service server (300) may quantify the semantic similarity of these multiple candidates based on the following multiple criteria. As a first criterion, the degree of contextual match with menu information may be determined. For example, this includes determining whether components such as "Yukgaejang = beef + bracken + spicy soup" exist in the candidate sentence. This degree of contextual match may be determined based on sentence embedding comparison based on cosine similarity, and, for example, a multilingual model derived from Sentence-BERT may be used. As a second criterion, whether the knowledge graph path matches may be determined. An example of this is determining whether a path stored in the knowledge graph (500), for example (Yukgaejang, includes, bracken) or (Yukgaejang, regional origin, Gyeongsang-do), is implicitly reflected in the natural language sentence in the form of "bracken included" or "Gyeongsang-do origin".This path determination can be performed using embedding-based triple mapping verification or rule-based keyword mapping methods. As a third criterion, it can be determined whether user settings are reflected. For example, if "Include allergy information" is selected according to user settings, bonus points are awarded only if the sentence includes phrases such as "possibility of causing beef protein allergies." According to the embodiment, the third criterion may be omitted. Subsequently, based on the scores calculated from the aforementioned plurality of criteria, a final score for semantic similarity can be calculated as shown in the following Equation 1. [Equation 1] Final_Score_i = w1·SimA_i + w2·SimB_i + w3·SimC_i Here, SimA_i: semantic similarity between candidate i and menu description (0~1) SimB_i: whether candidate i matches the knowledge graph path SimC_i: how much candidate i satisfies user-set conditions w1, w2, w3: weight parameters (e.g., 0.4, 0.4, 0.2). These similarity scores are compared as normalized values, and the translation candidate that obtains the highest score can be selected as the final result. In step S750, the service server (300) can finally adjust the stylistic style, expression length, emphasis order, etc., for the selected optimal translation result according to user settings. For example, - For example, when a formal style is requested, the text is corrected to end with the "desu" phrase; when allergy information is excluded, the corresponding item is removed; and when a summary model is called to concisely summarize the sentence length, such post-processing can be performed through a rule-based rewriting engine or an LLM-based post-processor module, and can apply template-based correction logic that incorporates sentence component analysis and language-specific grammar rules. An embodiment of FIG. 7 can not only generate high-quality translation results that reflect the semantic context based on a knowledge graph through such a series of procedures, but also provide diversified expression adjustment functions according to user settings.FIG. 8 is a block diagram illustrating the configuration of a restaurant information provision system based on multinational visitors according to an embodiment of the present invention. The restaurant information provision system based on multinational visitors relates to an embodiment that enhances the restaurant information provision function based on multinational visitors in the multilingual menu system described above with reference to FIG. 1 to 7. Referring to FIG. 8, a service server (301) may include a sales information collection unit (310), a sales information management unit (320), a multilingual processing model (330), an explanation information generation AI (340), an additional information generation AI (350), a visit information management unit (360), and a restaurant map management unit (370). The basic functions of the sales information collection unit (310), the sales information management unit (320), the multilingual processing model (330), the explanation information generation AI (340), and the additional information generation AI (350) can be easily understood from the description described above with reference to FIG. 1 to 7. The visit information management unit (360) together with a review upload request received from a user terminal The system performs a function to estimate the nationality of visitors based on the language setting information of the user terminal, browser locale information, IP address, etc., by analyzing the acquired request header data. The estimated nationality information generated at this time is stored in conjunction with the corresponding user review and can be used to analyze visitor statistics by nationality and preference patterns by cultural region. The restaurant map management unit (370) performs a function to visually organize map-based restaurant information using multinational review data. To this end, it visualizes store locations, review reactions, preferred menus, etc. through a restaurant map generation function and can filter out fake or duplicate reviews through a review verification function. In addition, it may include a function to provide recommendation information customized to users of a specific nationality by reflecting review trends by country or cultural region. The multilingual processing model (330) is a module for processing multilingual review data or generated food description information and includes a prompt generation function and a data set linkage function.The prompt generation function generates structured input including domain context for the sentence to be translated based on external multilingual knowledge data (500), enabling it to be linked with a large-scale language model (LLM) (400). The descriptive information generation AI (340) is an artificial intelligence module that automatically generates summary descriptions, service characteristics, popular menu items, etc., about a restaurant in the form of sentences, and is designed to automatically construct expressions that are easy for foreign visitors to understand. The descriptive information generated at this time can be translated into multiple languages ​​through a multilingual processing model (330) and an LLM (400) and provided to the user. With this configuration, in one embodiment of the present invention, beyond simple translation, restaurant information based on the reactions and statistics of multinational visitors can be generated, and customized recommendation information by country and cultural region can be provided. Furthermore, insights based on review characteristics and preferences by cultural region can be provided to store operators, thereby contributing to the improvement of operational efficiency and customer satisfaction in the food service industry. FIG. 9 is a flowchart illustrating a restaurant review processing procedure for estimating multinational visitor information according to an embodiment of the present invention. In step S901, a user terminal (200) can recognize unique store identification code information from a store code (e.g., QR code or NFC tag, etc.) placed in the store. This identification code information is a code uniquely assigned to the store and serves as a criterion for identifying which store subsequent review data is associated with. In step S902, a service server (301) receives identification code information from the user terminal (200), checks the corresponding store page, and provides a web-based user interface to the user terminal (200) that enables writing a review for the store. For example, the web-based user interface may consist of review writing items, star rating selection, photo upload, translation options, etc.In step S903, the user terminal (200) writes a review through a web-based user interface and sends a review upload request to the service server (301). At this time, the review upload request includes request header data containing the user's browser settings, language environment, locale settings, IP address, etc. In step S904, the service server (301) identifies and analyzes the request header data included in the review upload request. In one embodiment, the service server (301) extracts information such as language setting information, browser locale, and IP address to estimate which country or cultural region the user of the corresponding user terminal belongs to. For example, if the language setting is French and the IP address originates from the Jeonju region of the country, it is presumed to be a review by a French national user. In step S905, the service server (301) stores the user's estimated nationality information, generated based on the analyzed request header information, together with the uploaded review data. This stored information can subsequently be utilized for nationality-based statistical analysis, comparison of reactions by cultural region, and provision of customized restaurant recommendation information. Through this procedure, in one embodiment of the present invention, the nationality information of a visitor is effectively estimated using the technical setting information of the user terminal without separate membership registration or input of personal information, and by accumulating this information in conjunction with the review data, the effect of automatically collecting review reactions and visit statistics by country or cultural region is provided. FIG. 10 is a diagram illustrating an example of the configuration of request data received when uploading a review according to one embodiment of the present invention. As shown in FIG. 10, when a user terminal transmits a request to upload a review to a service server, the request includes request data, and this request data includes request header data. Request header data is information that is automatically included depending on the user's terminal environment, and includes the user's browser settings, language settings, locale, and other system information.In particular, in one embodiment of the present invention, the user's nationality or language culture is estimated using "user browser information" and "user language setting" items among the request header data. For example, user browser information may include a 'User-Agent field' and, as shown in the example below, includes information regarding the browser type, version, operating system, device type, etc. Field iPhone; CPU iPhone OS 16_5 like Mac OS X User language setting information may include, for example, an 'Accept-Language field', which refers to the language environment set by the user in the browser or operating system. This Accept-Language field may include the following items. Accept-Language: fr-FR,fr;q=0.9,en-US;q=0.8,en;q=0.7 In the above example, it means that the user browser has set French (fr-FR) as the default language, and English (en-US, en) is also set as an acceptable language. By analyzing such header information, the present invention can presume that the user is a French-speaking user and assume French nationality. Furthermore, the request data may include IP address-based connection location information, which enables additional determination of the country. For example, if the IP belongs to an IP range issued in France, it is possible to confirm the user's nationality as France by combining it with the aforementioned Accept-Language item. Thus, one embodiment of the present invention provides the effect of estimating the user's nationality by utilizing configuration information included in the request header data without the need for a separate login or user account creation, and by storing this information in association with review data, it can be utilized for purposes such as country-specific visit statistics and language-based restaurant analysis.In particular, the present invention has a technical advancement in that it enables more precise and reliable nationality estimation by comprehensively considering multiple factors, such as browser settings, compared to existing simple location-based estimation methods (using only IP addresses). FIG. 11 is a flowchart showing the processing sequence for estimating nationality or cultural sphere based on user language preference information according to an embodiment of the present invention. In step S1110, the service server (301) can extract the Accept-Language field included in the request header of the request upon receiving a review upload request from the user terminal (200). As described above, the Accept-Language field includes multiple language information assigned priority according to the user's browser or operating system settings. For example, it may be in the form of fr-FR;q=1.0, en-US;q=0.9, zu;q=0.8. In step S1120, the service server (301) can parse the extracted Accept-Language field to generate a language priority list. At this time, each language code and weight information (q value) can be separated into individual items and stored as a structured list. For example, it may take the form of [(fr-FR, 1.0), (en-US, 0.9), (zu, 0.8)]. In step S1130, the service server (301) can search for a list of associated countries based on language-specific weight scores and calculate a cumulative score for each country. Specifically, the service server (301) can search for multiple countries Cj corresponding to each language code Li in the parsed language priority list through a mapping table and obtain a ratio value rij for each country, which can be expressed as [Equation 2] below. [Equation 2] Here, Sj is the cumulative score of country Cj, wi is the priority weight (q-value) of language Li, and rij is the degree to which language Li is associated with country Cj (weight distribution ratio). For example, fr is France (0.5), Cameroon (0.3), Belgium (0.Let us assume that the data is divided and mapped as follows, with en mapped to the United States (0.6), the Philippines (0.25), Nigeria (0.15), and zu to South Africa (1.0). In this example, the service server (301) can calculate a cumulative score for each country as shown in the following [Equation 3]. [Equation 3] In step S1140, the service server (301) can estimate the country with the highest value among the cumulative scores Sj as the user's nationality. For example, in the above case, since South Africa has the highest cumulative score of 0.8 points, it can be determined that the user is likely of South African nationality. As such, the method according to one embodiment of FIG. 11 can improve the accuracy of identifying foreign users by implementing a complex nationality estimation logic that numerically reflects the distribution of language preferences and cultural overlap, going beyond single language information or simple location-based estimation methods. FIG. 12 is a block diagram illustrating the configuration of a web-based multilingual menu ordering system according to one embodiment of the present invention. The web-based multilingual menu ordering system relates to one embodiment that enhances the online ordering function of the multilingual menu system described above with reference to FIGS. 1 to 7. The web-based multilingual menu ordering system can provide a reliable ordering environment without the need for app installation by verifying whether a store code-based order request was made within the actual store and determining whether the order is valid based on the user's location information. Referring to FIG. 12, the service server (301) may include a sales information collection unit (310), a sales information management unit (320), a multilingual processing model (330), an explanation information generation AI (340), an additional information generation AI (350), and an order management unit (380). The basic functions of the sales information collection unit (310), the sales information management unit (320), the multilingual processing model (330), the explanation information generation AI (340), and the additional information generation AI (350) can be easily understood from the description above with reference to FIGS. 1 to 7.The order management unit (380) is a core component that performs a web-based multilingual menu ordering function capable of location-based verification in this embodiment. The order management unit (380) may largely include an order interface provision function and a location-based verification function. In one embodiment, the order management unit (380) receives store code information from a user terminal and, based on the received store code information, provides a web-based multilingual ordering interface for the corresponding store to the user terminal. For example, a QR code attached to a store table may include a unique store code or a table identification code. When a user scans a QR code, the code is transmitted to a server, and the order management unit (380) can interpret it and load a multilingual order UI dedicated to that store into the browser of the user terminal. The order interface can automatically detect the browser language setting information of the user terminal and automatically set the multilingual language to provide a web-based multilingual order interface. When an order request is received from a user terminal, the order management unit (380) can perform a procedure to verify whether the request was actually made within the store. In one embodiment, the order management unit (380) can obtain the latitude and longitude coordinates of the user terminal based on the GPS information of the user terminal and verify the reference location coordinates corresponding to the store code (e.g., verifying latitude and longitude coordinates based on the map location or GPS location of the store). Subsequently, a distance value between two locations is calculated based on the latitude and longitude coordinates of the user terminal and the latitude and longitude of the store, and the order can be verified as valid only if this value is less than or equal to a predefined allowable distance threshold. In one embodiment, the threshold may vary depending on the type of store, and may be dynamically set, for example, 20m for a standalone building, 100m for a store within a large shopping mall, etc.This is intended to set a small error range for stores in standalone buildings and a large error range for stores within large buildings, due to the current limitation that the error range of GPS location increases when inside a building. In one embodiment, the location of a user terminal can be estimated using internet-based location information instead of or together with GPS information. For example, connection time, IP address, and user agent information upon request for a store code can be collected, and verification can be performed based on whether the information originated near the store or within the internal network. In one embodiment, the store code information may include an identification code uniquely assigned to each table unit rather than a single store, and this table identification information may also be included and transmitted to the server upon request for an order. In this embodiment, the store operator can accurately identify which table an order originated from and perform order processing, and can utilize this information for table-specific order history, turnover analysis, and customer waiting time improvement. In one embodiment, the order management unit (380) can transmit order information regarding a verified order request to the seller terminal or the store system. Additionally, the seller terminal (100) may be connected to a wireless printer separately installed in the store, in which case, a web-based output interface may be triggered on the seller terminal (100) to support printing of order details using the wireless printer. In another embodiment, a seller app may be installed on the seller terminal (100), and in this case, order details may be provided to the seller app of the seller terminal (100). The seller app may provide notifications regarding order details and may automatically print order details by linking with a wireless printer or POS device separately installed in the store, allowing kitchen or counter staff to check them immediately. This process is structurally simplified and has technical advantages in that it allows for the configuration of a simple smart order environment using only orders and a wireless printer, without the need for separate POS integration.FIG. 13 is a flowchart illustrating the processing procedure of a web-based multilingual menu ordering method capable of location-based order verification according to an embodiment of the present invention. In step S1301, a user terminal (200) can recognize a code attached within a store (e.g., a QR code or an NFC tag) to obtain store code information and transmit it to a service server (300). According to an embodiment, this store code may include a table identification code subdivided by table, and the service server (302) can determine which table an order has been placed at based on the information. In step S1302, the service server (302) can verify the registration status and validity of the store corresponding to the received store code information, and then provide a web-based multilingual ordering interface for the store to the user terminal (200). In one embodiment, the service server (302) can automatically set an optimized multilingual UI based on the browser language setting information or location information of the user terminal, which can be understood by referring to the embodiments described above with reference to FIGS. 8 to 11. In step S1303, the user terminal (200) can select order items through the provided multilingual order interface and perform a pre-payment input procedure. At this time, the input order request may automatically include the user terminal's current location information (latitude and longitude coordinates). Alternatively, the user terminal's information may be collected through a separately allowed browser location sharing API, etc. In step S1304, the service server (302) may verify the validity of the order request by calculating the geographical correlation between the user terminal's location information included in the order request and the store's reference location information corresponding to the store code. In one embodiment, the user terminal's location information may identify the orderer's wide area information based on the user terminal's IP. For example, the orderer's location at the country or city level may be identified using GeoIP.If such a location differs from the location of the store—for example, if the country or city is different—it may be processed as an invalid order. In one embodiment, the distance value between the two locations may be calculated using a location-based distance calculation algorithm such as the Haversine formula. The service server (302) determines the order to be valid only if this distance value is below a pre-set allowable distance threshold. In one embodiment, the allowable distance threshold may be dynamically set based on the store's location type (indoor / outdoor), building size, etc. In one embodiment, the service server (302) may verify whether access to the store is possible by considering metadata such as the time of connection, IP address, and terminal environment (agent) in addition to location information. In one embodiment, the service server (302) may perform order verification based on the menu and amount. For example, it may verify whether an actual existing menu was ordered and whether the amount exactly matches the database stored in the store. If it is determined to be a discrepancy, it may be determined to be a malicious order, and a process may be performed to block the corresponding user terminal for a predetermined time (e.g., 30 minutes). In one embodiment, the service server (302) may additionally perform a process to determine whether the orderer is a human. For example, a web-based human recognition script such as Google reCAPTCHA may be applied to the order interface to block malicious order attacks by robots rather than humans in advance. In step S1305, if the service server (302) determines that the order information is valid, it transmits the order information to the seller terminal (100) in real time. In one embodiment, the seller terminal may be linked with a wireless printer or POS device in the store, and the order management unit may process to print a physical order receipt or order sheet based on the received order information.In one embodiment of FIG. 13, the store owner is provided with an automated ordering process without a separate app or login process, and the customer can also enjoy a convenient and reliable ordering service in a multilingual environment. FIG. 14 is a block diagram illustrating an order management unit that automatically analyzes abnormal ordering behavior of a user and presents correction candidates according to one embodiment of the present invention, and FIG. 15 is a flowchart explaining an order management method performed in the order management unit of FIG. 14. The order management unit (380) of FIG. 14 may further include a user behavior analyzer, an order correction recommender, and an allergy information filter in addition to the order interface, location-based verifier, and printer linker described above. The user behavior analyzer can analyze the user's ordering behavior information in real time and determine the possibility of an error. The order correction recommender can generate a correction candidate menu and present it to the user. The allergy information filter can identify dangerous menus and issue warnings or suggest alternative menus by referring to ingredient information included in the menu and allergy statistics by user country. Referring further to FIG. 15, in step S1510, user behavior information occurring when an order is executed by a user terminal of the order management unit (380) is collected. User behavior information may include, for example, information such as menu navigation patterns, selection delay times, menu change history, and past order history occurring on the user terminal. This user behavior information is obtained through interaction events of a web-based order interface or script-based logging techniques. In step S1520, the user behavior analyzer module of the order management unit (380) analyzes the collected behavior information to determine whether there is a possibility that the ordering behavior currently being performed by the user is an error. For example, if the selection trend of a menu repeatedly selected in the same language group is statistically significantly different from that of a different language group, it may be determined as an abnormal selection.Additionally, circumstances such as the user quickly deleting a specific menu immediately after selecting it, or selecting a menu containing allergy ingredients, can also serve as grounds for determining the possibility of an error. In step S1530, if the judgment result indicates that correction is required, the order correction recommendation module of the order management unit (380) automatically generates alternative correction candidates for the user's selected menu. For example, if the selected menu is a spicy dish containing chili ingredients, and the user's past order history or location-based country culture information indicates low consumption of spicy food, a similar non-spicy menu may be presented as a correction candidate. This process is carried out by referring to multilingual knowledge data (500) and explanatory information and additional information of the sales information management unit (320), and multilingual semantic similarity analysis through LLM (400) may be utilized. In step S1540, the generated correction candidates are presented on the order interface of the user terminal, and the user is encouraged to re-select one of the correction candidates. At this time, the re-selection may be provided as a user-friendly conversational sentence using LLM. For example, it can be output in a user-friendly sentence format such as, "This Menu A may conflict with allergy information. How about Menu B as an alternative menu?" An embodiment disclosed in FIGS. 14 and 15 provides an effect that effectively reduces the problem of misordering, particularly in a multilingual environment, by going beyond the existing simple order acceptance function and providing a smart ordering experience that interprets user behavior and minimizes cognitive errors. FIG. 16 is a block diagram showing the configuration of a multilingual question-and-answer providing system using an intelligent chatbot according to an embodiment of the present invention. The multilingual question-and-answer providing system using an intelligent chatbot relates to an embodiment that enhances the chatbot-based customer response function in the multilingual menu system described above with reference to FIGS. 1 to 7.The web-based multilingual menu ordering system can provide a reliable ordering environment without the need for app installation by verifying whether a store code-based order request was executed within the actual store and determining whether the order is valid based on the user's location information. Referring to FIG. 12, the service server (301) may include a sales information collection unit (310), a sales information management unit (320), a multilingual processing model (330), an explanation information generation AI (340), an additional information generation AI (350), and an intelligent chatbot (390). The basic functions of the sales information collection unit (310), the sales information management unit (320), the multilingual processing model (330), the explanation information generation AI (340), and the additional information generation AI (350) can be easily understood from the description above with reference to FIGS. 1 to 7. The intelligent chatbot (390) is responsible for the function of interactively receiving user queries and generating appropriate responses accordingly. The intelligent chatbot (390) includes a chatbot interface module, an intelligent response processing module, and a HITL (Human In The Loop) processing module. The chatbot interface is called via a chatbot connection element (e.g., button, banner, etc.) included in a web-based multilingual menu interface on a user terminal and receives user queries in real time. Users can input various menu-related queries in natural language format. When a user query is received, the intelligent response processing module first refers to chatbot QnA data (391) to determine whether a suitable response exists. At this time, keywords are extracted from the user query, and it is checked whether question-response pairs similar to the keywords exist in the QnA database. If a suitable response is found, the response data is provided to the user.The Human In The Loop (HITL) processing module performs a Human In The Loop (HITL) method of generating a response by sending the query to an expert when a suitable response is not found in the chatbot QnA data (391) or when the confidence score of the response data generated through the large-scale language model (400) is determined to be below a threshold. For example, if the confidence score of the response is 0.5 or lower, or if a NULL response is derived, the system automatically activates the HITL path. In the HITL method, expert information capable of responding to the user query is linked with a pre-registered expert mapping database, and a subject manager similar to the query is selected to directly connect to the user's response environment via phone, email, or real-time chat. The descriptive information generation AI (340) and the supplementary information generation AI (350) can be linked with the large-scale language model (400) and used as an auxiliary tool when constructing response sentences. The generated response data and expert response history are stored in the chatbot QnA data (391) and can be reused as preliminary responses for similar queries in the future. This method improves the self-learning performance of the system and has the effect of improving the accuracy and speed of responses to repeated queries. In addition, the intelligent chatbot (390) can optimize and provide the response language according to the language setting information of the user terminal, thereby providing the user with a natural and intuitive multilingual question-and-answer environment. As such, according to the present embodiment, by implementing a hybrid response structure that combines existing QnA data-based responses, LLM-based responses, and expert responses in stages for a user's query, it is possible to provide a question-and-answer environment with high accuracy and high reliability. FIG. 17 is a flowchart showing the processing flow of a method for providing multilingual question-and-answer using an intelligent chatbot according to an embodiment of the present invention.This embodiment is performed through the cooperation of a user terminal (200), a service server (303), and a large-scale language model (400). In step S1701, the user terminal (200) transmits a conversational user query through a multilingual menu chatbot interface. The chatbot interface is a component embedded in a web-based multilingual menu UI, and the interface is displayed according to the language the user wishes to query, and can be activated through operations such as clicking or touching. In one embodiment, the chatbot interface can be accessed through a multilingual menu service interface. For example, when the user terminal provides store code information to the service server (303) using the store code of a store, the service server (303) provides a web-based multilingual menu interface for the corresponding store to the user terminal as described above. When a user terminal (200) clicks a chatbot connection element (e.g., a banner or a specific button, etc.) included in the web-based multilingual menu interface of the store, the service server (303) can provide the web-based multilingual menu chatbot interface to the user terminal (200). In this case, the default language can be automatically switched according to the user's terminal language settings, thereby enabling a natural multilingual chatbot experience. In step S1702, the service server (303) extracts keywords included in the user query and, based on this, combines the chatbot QnA database and the large-scale language model (400) to generate response data for the query. For example, the service server (303) first performs a keyword-based database search, and if suitable response data exists, it utilizes it as is or provides supplementary contextual information to the large-scale language model to generate a natural conversational response sentence for the user query. In one embodiment, if no suitable response exists in the QnA database, the service server (303) calls the large-scale language model alone to respond to the query You can generate a direct response to [the].The generated response data is scored according to internal reliability evaluation criteria, and if this reliability is determined to be below a preset threshold, the process proceeds to step S1703. In step S1703, the service server (303) performs a Human In the Loop (HITL) procedure to select expert personnel associated with the user query and have them directly participate in the response environment. For example, if the quality score of the response is 0.4 or lower after the response is generated, the system may be configured to automatically identify a person in charge of the relevant domain from a pool of experts registered in advance, connect them to the real-time response environment, or send a follow-up response notification. In one embodiment, the system may be configured so that an expert using an appropriate language is selected to respond based on the user's receiving language settings. In one embodiment, the final generated response sentence and the expert intervention history accompanying the response are all stored in the chatbot QnA database and may be configured to be reused as prior response data when the same or similar query occurs in the future. Through such a data storage and retraining structure, the system can realize a self-evolving structure that gradually improves response accuracy and processing efficiency. As such, in this embodiment, by combining the existing structured Q&A-based response method, the natural language generation capability of a large-scale language model, and a HITL structure that allows human intervention depending on the situation, a highly reliable multilingual question-and-answer environment can be realized, and in particular, issues regarding response quality and user distrust in actual service environments can be effectively resolved. FIG. 18 is a flowchart illustrating a method for generating conversational response data for a user query according to an embodiment of the present invention. In step S1810, the service server receives a conversational user query received through a user terminal.In this step, to process user queries that may be entered in multiple languages, in one embodiment, a language detection function may be performed in parallel to identify the input language of the query or to split and normalize mixed-language sentences by language. This can improve the accuracy of subsequent query analysis. In step S1820, the service server uses a Natural Language Processing (NLP) module to extract major noun phrases, named entities, or intentions from the user query and identifies them as keywords. In one embodiment, this NLP processing process may be configured to accurately identify semantic differences based on context by utilizing a parser and tokenizer optimized for each language according to the multilingual detection results. In addition, sentiment analysis is performed in parallel to identify the emotional nuances of the query, and the response tone can be adjusted based on this. In step S1830, an internal chatbot QnA database is queried based on the extracted keywords to collect a set of response candidates capable of responding to the user query. For example, this step is based on a pre-configured similar query and response dataset, and in one embodiment, similarity-based response candidates can be selected as Top-N using methods such as BM25 or Dense Retrieval (BERT-based). Additionally, context-based response candidate correction can be performed by considering the recent user conversation context. In step S1840, a structured prompt is generated including the collected response candidates, keywords, and the entire content of the user query, and is provided to a Large Language Model (LLM). Unlike existing methods that provide only simple query sentences when configuring a prompt, the present invention provides an advanced effect by including the following information together.- Key keywords and semantic tags extracted from the query (e.g., food name, allergy information, recipe) - Existing Q&A responses to similar queries - Information on relationships between entities inferred from the knowledge graph (e.g., 'kimchi', 'fermented food', 'spicy level') - Cultural style guides based on user terminal language / location / country. This configuration induces the LLM to generate more accurate and natural multilingual responses, which is one of the key technical differentiators of the present invention. In step S1850, among the multiple response candidates generated by the LLM, filtering is performed based on semantic consistency, stylistic appropriateness, tone of response, potential for causing offense, and whether translation is omitted. In one embodiment, a quality score is calculated based on BLEU, ROUGE, or embedding-based semantic similarity measurements for the response candidates, and responses below a threshold are automatically removed. This process serves as a pre-cleaning filter to reduce HITL resource consumption in subsequent steps. In step S1860, the quality reliability of the filtered responses is re-evaluated, and it is determined whether the response requires HITL (Human-In-The-Loop) review. In one embodiment, a HITL trigger is activated if one or more of the following conditions are satisfied: - Missing or conflicting named entities in the response (e.g., mention of pork in pork-free dishes) - Detection of syllable / grammar errors - Discrepancy in the basis of response information on the knowledge graph - When the sentiment analysis result is negative or excessively positive rather than neutral. In step S1870, if the HITL trigger conditions are satisfied, suitable experts (e.g., food experts, foreign language translators, etc.) are matched based on user query and response candidate information, and they participate in the generation of HITL response data. In one embodiment, the HITL participant performs supplementation, modification, approval, etc., by referring to the context of the query and response, and this process can be managed through a separate HITL UI.Through these steps, the present invention can provide accurate and highly reliable multilingual conversational responses to user queries by including not just simple LLM response generation, but semantic-based multi-stage filtering, reliability evaluation, and an expert intervention-based supplementation process. FIG. 19 is a flowchart illustrating an expert pool classification and dynamic selection method for selecting appropriate Human-In-The-Loop (HITL) experts according to the reliability and characteristics of a user query, according to an embodiment of the present invention. This embodiment includes a series of processes to improve response quality by classifying experts according to the context of the user query and the response risk level, and automatically selecting the HITL expert most suitable for the query. In step S1910, the service server (303) configures a plurality of HITL expert pools according to predefined criteria to respond to user queries that may occur in various language / domain / cultural contexts. For example, the expert pools may be classified into the following multi-layered structure: - Language correction expert pool: Review of translation fluency and stylistic consistency - Food domain expert pool: Food names, Experts in food information such as ingredients, allergens, and cooking methods - User feedback-based UX expert pool: analyzes user responses to similar past queries and reconstructs more user-friendly responses - Regulatory sensitivity expert pool: responsible for reviewing sensitive items such as local food regulations and health notices. A structure like this provides a foundation for personnel with different expertise to intervene depending on the attributes and risk level of the user query. In step S1910, the characteristics of the user query and the risk level are evaluated. That is, the service server (303) receives a query entered from the user terminal (200) and evaluates the risk level (RL) based on the subject, sentence structure, included entities, cultural / regional elements, and required level of explanation of the query. For example, the risk can be quantified based on the following criteria.For example, if allergens are mentioned in the query: RL = High; if subjective elements such as spiciness level or taste description are included in the query: RL = Medium; if the query corresponds to a simple information request: RL = Low. These evaluation results are used as criteria for expert selection in subsequent steps. In step S1930, expert group matching based on query attributes is performed. That is, the service server (303) extracts expert groups requiring selection based on the risk level and query topic tag calculated in step S1920. For example, RL=High, Topic=Allergy: "Food Domain Expert + Regulation Sensitivity Expert"; RL=Medium, Topic=Spiciness: "Food Domain Expert + UX Expert"; RL=Low, Topic=Ingredient Description: "Language Correction Expert". In this way, it is designed so that multiple expert groups can intervene jointly or a single expert can be automatically matched. In step S1940, candidate alignment based on expert trust scores is performed. That is, the service server (303) selects an expert to select an appropriate respondent from among multiple experts within the same expert group. A matching score (HMS: HITL Matching Score) is calculated. In one embodiment, the expert matching score can be calculated using a weight-based formula as shown in [Equation 4] below. [Equation 4] Here, A = Accuracy Score: Average user evaluation of past responses (0~1), F = Feedback Suitability Score: User satisfaction with similar queries, R = Response Latency Score: Reciprocal normalization of average response time (higher the faster), D = Domain Expertise Score: Quantified knowledge level regarding the subject (test score, qualification, etc.), L = Language Appropriateness Score: Language consistency evaluation score, w1 ~ w5: Weights assigned to each element, for example (0.3, 0.2, 0.2, 0.2, 0.1).For example, if Expert A has scores of A = 0.95, F = 0.90, R = 0.85, D = 0.80, L = 0.75 and Expert B has scores of A = 0.97, F = 0.85, R = 0.60, D = 0.90, L = 0.80, then Expert A's expert matching score is calculated as = 0.3 x 0.95 + 0.2 x 0.90 + 0.2 x 0.85 + 0.2 x 0.80 + 0.1 x 0.75 = 0.875, and Expert B's expert matching score is calculated as 0.3 x 0.97 + 0.2 x 0.85 + 0.2 x 0.60 + 0.2 x 0.90 + 0.1 x 0.80 = 0.843, resulting in Expert A being matched first. This HMS is continuously It is updated and may reflect the time-based weighted average of recent responses, thereby enabling dynamic and highly real-time expert sorting. Subsequently, in step S1950, real-time availability judgment and final expert assignment are performed. The service server (303) selects the final responding experts based on real-time availability information from the experts in the expert list sorted based on the HITL matching score (HMS) calculated in the previous step. In one embodiment, for this purpose, the expert profile used for real-time availability information may include the following information: current login / session status, response time zone and priority settings, number of concurrent responses processed (Concurrent Task Limit), and past fatigue indicators (e.g., number of responses within the last hour). Based on this expert profile information, the service server (303) may finally select an expert based on the following real-time availability filtering algorithm. Only experts whose HMS is within the top 10% are set as candidates; among these candidates, only experts who are currently 'online', whose concurrent processing volume is below a threshold, and whose recent feedback rating has not dropped are selected. After this multi-stage filtering, a final one is assigned, or multiple if necessary. Personnel (e.g., selection of post-processing after multiple responses) can be assigned in parallel.Subsequently, in step S1960, HITL response intervention and result storage are performed. The service server (303) provides the user query and intermediate response data generated by LLM to the finally selected HITL expert, and the HITL expert performs an overall review of the LLM response, such as accuracy, fluency, tone, and cultural appropriateness; directly modifies or supplements the response if necessary; and reconstructs the response based on a portion of the LLM-generated response. The response generated by the expert may be stored along with meta-information, such as HITL ID, response time, ratio of changed sentences, quality score, and question-response pairs as retraining data, for future learning purposes. This embodiment provides the effect of increasing the efficiency of expert allocation by considering real-time response processing status and fatigue information, and enabling the realization of compatibility between system automation and response quality control as all of this processing is operated by an automated selection algorithm. FIG. 20 is a diagram illustrating an exemplary computing operating environment of a service server according to an embodiment of the present invention. FIG. 20 is a suitable embodiment in which embodiments of the service server (300) can be implemented. Referring to FIG. 20, for the purpose of providing a general and simplified description of a computing environment, a computing device is illustrated as an example of a service server (300, 301, 302, 303). The computing device may include at least a processing unit (303) and a system memory (301). The computing device may include a plurality of processing units that cooperate when executing a program. Depending on the exact configuration and type of the computing device, the system memory (301) may be volatile (e.g., RAM), non-volatile (e.g., ROM, flash memory, etc.), or a combination thereof. The system memory (301) includes a suitable operating system (302) for controlling the operation of the platform, which may be, for example, the WINDOWS operating system from Microsoft.System memory (301) may include one or more software applications, such as program modules, applications, etc. The computing device may include additional data storage devices (304), such as magnetic disks, optical disks, or tapes. These additional storage devices may be removable storage and / or fixed storage. Computer-readable storage media may include volatile and non-volatile, removable and fixed media implemented by any method or technique for storing information such as computer-readable instructions, data structures, program modules, or other data. System memory (301) and storage (304) are all merely examples of computer-readable storage media. A computer-readable storage medium may include, but is not limited to, RAM, ROM, EERPOM, flash memory or other memory techniques, CD-ROM, DVD or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage device, or any other medium that can store desired information and be accessed by a computing device (300). An input device (305) of the computing device may include, for example, a keyboard, a mouse, a pen, a voice input device, a touch input device, and a comparable input device. An output device (306) may include, for example, a display, a speaker, a printer, and other types of output devices. Since these devices are widely known in the art, a detailed description is omitted. The computing device may include a communication device (307) that allows the device to communicate with other devices through a network in a distributed computing environment, for example, a wired / wireless network, a satellite link, a cellular link, a local area network, and a comparable mechanism. A communication device (307) is one example of a communication medium, and the communication medium may contain computer-readable instructions, data structures, program modules, or other data.For example, communication media include, but are not limited to, wired media such as wired networks or direct wired connections, and wireless media such as acoustic, RF, infrared, and other wireless media. Methods according to various embodiments of the present application may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiments, or may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware device described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa. Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above.For example, appropriate results may be achieved even if the described techniques are performed in a different order than the described method, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from the described method, or are replaced or substituted by other components or equivalents. Therefore, other implementations, other embodiments, and equivalents to the claims are also included within the scope of the claims set forth below. Explanation of symbols 100: User terminal 200: Customer terminal 300: Service server 310: Sales information collection unit 320: Sales information management unit 330: Multilingual processing model 340: Description information generation AI 350: Additional information generation AI 380: Order management unit 301: System memory 302: Operating system 303: Processing unit 304: Storage 305: Input device 306: Output device 307: Communication device.

Claims

Claim 1 A method for providing multilingual menu translation performed in a multilingual menu translation system comprises: a step of confirming a user request including original menu information to be translated; a step of searching for domain information related to the original menu information in knowledge data based on the original menu information; a step of generating a prompt for translating the original menu information based on the searched knowledge data; and a step of inputting the generated prompt into a large-scale language model to generate a multilingual translation result; wherein the step of searching for domain information related to the original menu information in knowledge data includes a step of extracting keywords based on food names included in the original menu information and searching for related information including the keywords in knowledge data based on content; and wherein the knowledge data comprises a plurality of entities classified into cooking entities corresponding to food names, ingredient entities corresponding to ingredients, and origin entities corresponding to regions; A method for providing multilingual menu translation by fusing knowledge data and a language model, comprising: a knowledge graph that expresses the semantic relationships between the plurality of entities in a triple structure; wherein the knowledge graph provides association information for translation by exploring related relationship paths based on entities associated with keywords; and the step of searching for association information including keywords in knowledge data based on content includes: a step of expanding the search for entities including expressions or synonyms similar to the keywords; a step of evaluating multiple paths based on relationship attributes between the expanded entities and other entities; and a step of setting the association information by preferentially selecting the path with the highest association. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 A method for providing multilingual menu translation by fusing knowledge data and a language model, wherein, in claim 1, the step of setting the association information by preferentially selecting the path with the highest association includes the step of setting the path containing the material object or region object most closely connected to the food name as a priority evaluation target. Claim 6 A method for providing multilingual menu translation by fusing knowledge data and a language model, wherein the step of generating the prompt comprises: generating a food description sentence by reflecting at least one of a food name, ingredients, a recipe, and regional characteristics, including entity information extracted from the searched domain information; and generating a structured prompt including a multilingual translation target language, translation tone, and stylistic information together with the food description sentence. Claim 7 A method for providing multilingual menu translation by fusing knowledge data and a language model, wherein the step of generating the multilingual translation result comprises: a step of identifying a plurality of translation candidates generated from a language model based on the prompt; and a step of calculating the semantic similarity between the food name to be translated and the knowledge data for each translation candidate, and selecting the translation candidate with the highest semantic similarity as the final translation result. Claim 8 A method for providing multilingual menu translation by fusing knowledge data and a language model, wherein the step of generating the multilingual translation result comprises: a step of determining user characteristics based on the user's language settings, location information, or terminal information from the user terminal; and a step of generating a translation result by dynamically setting a translation style, explanation level, and output item according to the user characteristics. Claim 9 A seller terminal providing a seller request including original menu information to be translated; and a service server that, based on the original menu information, searches for domain information related to the original menu information in knowledge data, generates a prompt for translating the original menu information based on the searched knowledge data, and inputs the generated prompt into a large-scale language model to generate a multilingual translation result; A multilingual menu translation system that fuses knowledge data and a language model, comprising: a service server that extracts keywords based on food names included in the original menu information and searches for related information containing the keywords in the knowledge data based on content to search for domain information related to the original menu information in the knowledge data; the knowledge data comprises a plurality of entities classified into cooking entities corresponding to food names, ingredient entities corresponding to ingredients, and origin entities corresponding to regions, and a knowledge graph that expresses the semantic relationships between the plurality of entities in a triple structure; the knowledge graph provides the related information for translation by searching for related relationship paths based on entities associated with the keywords; and the service server that, in order to search for related information containing the keywords in the knowledge data based on content, expands the search for entities containing expressions or synonyms similar to the extracted keywords, evaluates multiple paths based on relationship attributes between the expanded entities and other entities, and sets the related information by preferentially selecting the path with the highest correlation. Claim 10 A computer-readable recording medium having a computer program stored therein for performing a method of providing multilingual menu translation by fusing knowledge data and a language model according to any one of claims 1, 5 to 8, combined with hardware.

Images